built-in-nodes/EmptyLatentImage

[giscus[bot]]

built-in-nodes/EmptyLatentImage

The EmptyLatentImage node is designed to generate a blank latent space representation with specified dimensions and batch size.

https://docs.comfy.org/built-in-nodes/EmptyLatentImage

[Nuos]

empty-latent-image/图像潜空间，具有宽度、高度、批次属性，它的空间表示是怎样的，是tensor张量（结构化的数据）吗？ 所谓潜空间/latent-space 如何理解和表示，是一个数据结构空间吗？

[Nuos]

关于潜空间的操作、运算有哪些？KSampler

[lin-bot23]

Hi @Nuos, good questions. Yes — the latent is a real tensor, not an abstract data structure. In ComfyUI the EmptyLatentImage node allocates an all-zeros tensor of shape [batch_size, 4, height / 8, width / 8] (see the source: nodes.py EmptyLatentImage.generate — https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py), i.e. one 4-channel map at 1/8 the pixel resolution per image in the batch, created on the intermediate compute device.

On how to think about "latent space": it is not a special data structure or a separate space you browse — it is simply the coordinate space of that 4-channel tensor produced by the VAE encoder. Each position holds learned features that the diffusion model operates on directly, instead of pixel RGB values. The VAE decoder maps this representation back to a full-resolution RGB image. The exact shape (4 channels, 8x downscale) applies to standard SD-style latents; other model families may use different channel counts or compression ratios.

The operations performed on the latent are what the later nodes do with that tensor: KSampler adds noise and progressively denoises it toward the target distribution; VAE Encode/Decode move data between latent and pixel representations; and the Latent-* nodes (Upscale, FromBatch, AddNoise, etc.) manipulate the tensor directly.

The doc page currently only describes the node inputs/outputs at a high level. If you think it should include a short explanation of the underlying tensor shape and what "latent space" means, feel free to say so — happy to pass that on as a doc improvement suggestion.